Fluctuations in the Hopfield model at the critical temperature
Citation for published version (APA):
Gentz, B., & Löwe, M. (1998). Fluctuations in the Hopfield model at the critical temperature. (Report Eurandom; Vol. 98003). Technische Universiteit Eindhoven.
Document status and date: Published: 01/01/1998 Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers) Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
Report 98-003
Fluctuations in the Hopfield Model at the Critical Temperature
B. Gentz M. Lowe
AT THE CRITICAL TEMPERATURE
BARBARA GENTZ AND MATTHIAS LOWE
ABSTRACT. We investigate the fluctuations of the order parameter in the Hopfield model of spin glasses and neural networks at the critical temperature 11f3c = 1. The number of patternsM(N) is allowed to grow with the number N of spins but
the growth rate is subject to the constraintM(N)15 IN ~ O. As the system sizeN
increases, on a set of large probability the distribution of the appropriately scaled order parameter under the Gibbs measure comes arbitrarily close (in a metric which generates the weak topology) to a non-Gaussian measure which depends on the realization of the random patterns. This random measure is given explicitly by its (random) density.
1. INTRODUCTION
In 1977, Pastur and Figotin introduced and discussed a disordered version of the Curie-Weiss model of ferromagnets (see [29], [30]). Later their model became popular under the name Hopfield model because of its impact on the theory of neural networks achieved by its rediscovery and reinterpretation by Hopfield [21]. This versatility of the Hopfield model-namely that it can be regarded as a very simple model of the brain on one hand, and as a so-called spin glass (i. e., a disordered spin system) on the other hand-has been the driving force for its popularity and the efforts which have been undertaken to obtain a better understanding of the model. The neural network point of view has been taken in the original paper by Hopfield [21] for instance, as well as in the papers [27], [28], [23], [25], [26], and many others while in the seminal paper [29], as well as in [7], [8], [9], [3], [16], [17], [4], [5], and [31] the statistical-mechanics and thus the spin-glass aspect of the model have been in the centre of interest. Of course, it would be very difficult to give a complete list of all important papers in this area. For an overview of recent results on the Hopfield model and related models and results which deeply influenced our understanding of the model and even were able to justify some of the physicists' predictions (see [1], e. g.) we refer the reader to [31] and [11] and, in particular, [6] therein.
To be more specific, let us now define the Hopfield model. First of all we choose two numbers N, MEN which will denote the number of spins or "neurons" and the number of so-called patterns, respectively. In contrast to a previous paper [20], we shall now treat the case where
M
=M(N)
may depend onN.
Henceforth, we shall write M and thus drop its dependency on N whenever there is no danger of confusion and we shall refer explicitly to this dependency only when necessary. TheDate: December 11, 1998.
1991 Mathematics Subject Classification. 60F05, 60K35 (primary), 82C32 (secondary).
Key words and phrases. Hopfield model, spin glasses, neural networks, random disorder, limit theorems, non-Gaussian fluctuations, critical temperature.
random function
(1.1)
(1.3) denotes the Hamiltonian of the Hopfield model, which is a function of the spin configuration CT E
{-I,
+
I}N. The strength of the pair interaction is random as the variables ~r E {-I, +1} with ~r denoting the ith component of the tlth pattern are random. In this paper we shall assume that the ~r are i.i.d. unbiased random variables, i. e., that at given system size N, the family of random variables {~r:
i E{I, ...
,N},
tl E {I, ...,M(N)} }
is independent with1
lP(~r = +1) = lP(~r = -1) =
2"
for all i and tl. Expectations with respect to lP will be denoted by lEo Whenever convenient, we shall write ~ for the (N x M)-matrix consisting of the (~rkJl'while
~i =
(a, ...
,~f1) and ~Jl = (~r,...
,e/v),
respectively, stand for the ith row and the tlth column of this matrix, respectively.The spin variables are assumed to be independent with an unbiased a priori dis-tribution P, i. e.,
1
P(CTi = +1) = P(CTi = -1) =
-2
for all i E N. In addition, we shall assume throughout this paper that the family
{~r
:
i E {I, ... ,N},
tl E {I, ... ,M} }
is independent of the family of the spin variables {CTi :i E {I, ...,N} }.
The Hopfield model at temperature 1//3
E (0,00)
may now be identified with the Gibbs measure with respect to the Hamiltonian (1.1), i. e.,(2N,/3(CT)
=
2-Nexp{-/3HN(CT)}/ZN,/3, CT E {-I, +1}N, (1.2)where the so-called partition function
1
ZN,/3
=
2N2:=
exp{-/3HN(CT)}<TE{-1,+1}N
is the normalization which makes (2N,/3 a probability measure.
In order to understand the introduction of the order parameter in the Hopfield model note that the Hamiltonian (1.1) may be rewritten in the following convenient form as a quadratic functional of the so-called overlap mN:
. where (1.4) mN(CT) = (mJlN(CT))Jl=I,...,M with N mjACT)
=
2:=
~rCTi'
i=1(1.5)
Here and below, 11·112 denotes the Euclidean norm in }RM. The tlth component mjy
of the overlap mN compares the spin configuration to the tlth pattern ~Jl in such a way that a large absolute value ofmjy(CT) means that the spin configurationCTlargely agrees with ~Jl (or its negative). These configurations are of low energy according to (1.4). Therefore, the overlap is an important quantity for the investigation of the
Hopfield model, a so-called order parameter. Its distribution under f2N,{3 has been
of major interest in the study of the model and also will be central in this paper. In [7]' Bovier, Gayrard, and Picco established a law of large numbers for the distribution of the overlap under the Gibbs measure (}N,{3 which holds for JP>-almost all
realizations of the random patterns~. They showed that, whenever
M(N)/N
-+ 0, for JP>-almost all~, the distribution of the overlapmN
under the Gibbs measure with external magnetic field of strength h =j:. 0 in the direction of the first unit vector el of the canonical basis in }RM converges weakly towards the Dirac measure D±z({3)elconcentrated in
±z(,B)el
as first the system sizeN
-+ 00 and then the strengthh -+
o±.
Herez(,B)
denotes the largest rootz
E [0,1) of the Curie-Weiss equationz = tanh(,Bz).
Note that
z(,B)
=
0 for,B
~,Be
=
1, so that Do is the uni.que limiting measure in the high-temperature region,B
~,Be
= 1, whereasz(,B)
>
0 for,B
>
,Be,
so that in this regime there is no unique limiting point.Note that this result strongly resembles the law of large numbers for the mean magnetization in the Curie-Weiss model, see [14, Theorem IV.4.1(a)], for example. As already explained at the beginning this is, of course, not accidental, as the Hopfield model can be considered as a disordered version of the Curie-Weiss model and, indeed, for M = 1 the Hopfield model and the Curie-Weiss model agree by a simple "gauge transformation"
(i.
e., replacing (J'i by (J'i~l).On the scale of fluctuations, when analyzing the distribution of
VN(mN-z(,B)el)'
the character of the disorder becomes visible. Indeed, for M / N -+ 0 and
(,B,
h) =j:.(1,0), the overlap satisfies JP>-almost surely a central limit theorem with the covari-ance matrix which could be expected from the analogy with the Curie-Weiss model and a centring which differs in the cases
,B
>
1 or h =j:. 0 from the naively expected one by a ~-dependent adjustment, see [16], [17], [19] and Bovier and Gayrard [4].As shown in a previous paper [20], the influence of the disorder is even stronger when investigating the fluctuations of the overlap at the critical temperature
1/,B
=l/,Be
=
1, even whenM(N)
remains bounded. Recall that in the Curie-Weiss model the criticality at temperature1/,B
= 1 can also be seen as the breakdown of the central limit theorem. As a matter of fact at the critical temperature the magnetization in the Curie-Weiss model-scaled by a factorN1/4-converges weaklytowards a random variable given by its density with respect to Lebesgue measure which is proportional to exp( -x4/12), d. [14, Theorem V.9.5]. In [20] we showed that in the Hopfield model with finitely many patterns
(i.
e., with M not depending on N) the distribution of the overlap-scaled by the same factor N1/4-regarded
as a random variable
QN
taking values in the Polish space M1(}RM) of probabilitymeasures on}RM converges weakly (with respect to JP» to a limiting random measure
QM·
This limiting random measureQM
is given by its (random) density with respect to the M-dimensional Lebesgue measure which is proportional toexp ( - 112
f>; -
~
L
x;x~
+
L
~",VX"xv),
(1.6)J.!=l l~J.!<v~M l~J.!<v~M
where 1] is an M(M - 1)/2-dimensional Gaussian random variable with mean zero
and the covariance matrix being the identity matrix, namely, ~ = (~(J.!,v),(J.!',V') ) (J.!,V),(J.!' ,v')
and
E
_{I,
if(/1,
v) =(/1',
v'),(f.L,V),(f.LI,V' ) - 0
,
otherwise,
for 1 :::;
/1
<
v :::; M and 1 :::;/1'
<
v' :::; M.This shows that even for finite M at the critical temperature
1/13=
1, the fluctua-tions of the overlap depend strongly on the random disorder as even the distribution of the limiting fluctuations is random. Even to formulate the corresponding result for the case where the number of patternsM(N)
is actually growing withN
seemed to be difficult, since, on one hand, we don't have an "infinite-dimensional Lebesgue measure" as reference measure and, on the other hand, we cannot work with finite-dimensional projections (as in the Central Limit Theorem) either, since the "mixed terms" 2.:1:S;f.L<v:S;M TJf.L,Vxf.Lxv tend to "glue" together the coordinates.In this paper we circumvent these difficulties by not stating a limit theorem but by showing instead that the distance between the distribution QN of the scaled overlap
and the random measure QM becomes small with high probability for largeN. More precisely, we shall show, under the constraint M15/N -+ 0 on the growth rate of
M(N),
that for each large enoughN
there exists a set of ~'s of probability larger than 1-exp{-M/L} (with some constant L>
0) on which the distance betweenQN
and QM is smaller than eN "'" O.This paper has three more sections. Section 2 contains the explicit statement of the result concerning the non-Gaussian fluctuations of the overlap at
13
=
1 for the Hopfield model with a growing number of patterns. Section 3 is devoted to one of our basic tools, a multidimensional version of a strong approximation result of Koml6s, Major and Tusnady [22], which allows to control the difference of a sum of i.i.d. random variables and a sum of i.i.d. Gaussian random variables with the same covariance matrix. These results go back to Zaitsev [32], [33], Einmahl [12] and Einmahl and Mason [13]. They also proved useful in [10]. Section 4 finally is devoted to the proof which is based on the Hubbard-Stratonovich transform of the measures of interest together with a Taylor expansion of the resulting density, a saddle point approximation as well as the strong Gaussian 'approximation mentioned before.Acknowledgement. We are grateful to Anton Bovier for bringing the strong Gaussian approximation to our attention, and, in particular, for sharing the re-sults of [10] with us prior to publication. We benefited from interesting discussions with him. The results presented here were obtained while the second author was visiting at the WIAS. He thanks the WIAS for its hospitality.
2. STATEMENT OF RESULTS
This section contains the mathematically precise statement of the result an-nounced in the introduction. We shall state the theorem only for the case of
13
=13c
= 1 being fixed. In [20]' where we consideredM
independent ofN only, we also treated the case of variable temperature13N
converging to13c
=
1 as N -+ 00.It turned out that for
13N
converging to13c
faster thanI/VN
(recall that M was chosen as a constant), the limiting distribution is the same, while for13N
converg-ing to13c
slower thanI/VN,
we have a Central-Limit-Theorem type result and at "the borderline", i. e., when13N - 13c
is of the same order asI/VN,
one can see the influence of both possible limiting distributions.In the present setting, we consider such an extension of our results to variable
I3N
a basically technical exercise. Therefore, we shall concentrate on the most interesting case which allows us to present streamlined proofs.In general, we shall assume that the pattern matrix ~ lives on a probability space
(D,F, P) that is rich enough to allow the strong-approximation results stated in
Section 3. The pattern matrix has to be viewed as a random variable on (D, F, P),
but with slight abuse of notation, we shall formulate exceptional sets as sets of ~
variables by writing {~ :F(O E A} which is to be understood in the natural way as
{w
ED: F(~(w))EA}.
Let
QN = (IN,1(N1/4mN)-1 (2.1)
denote the distribution of the scaled overlap under the Gibbs measure (IN,1' By d
we denote the metric
d(P1 ,P2) = sup {
If
fdg -f
fdP21 : f Eg}
(2.2)with
9
=
(2.3){f: RM -+ R: sup If(x) - f(y)1 ::; 1 and sup If(x) - f(y)1 ::;
Ilx -
y112}
x,yEIRM x,yEIRM
on the set M 1(RM ) of all probability measures on RM. According to [2,
Corol-lary 2.8] this metric generates the weak topology on M 1(RM ). The result we are
going to prove is the following.
Theorem 2.1. Let
13
=
I3c =
1. Assume thatM(N)15 IN
-+ O. Then there exist aconstant L
>
0, a set D(N)c
D with probabilitylP(D(N))
2:
1 - e-M(N)/L, (2.4)an N E N and a sequence (EN )NEN, satisfying EN '\. 0 as N -+ 00, such that for
every N
2:
N, there exists a set(TJjl,Vh~/.l<v~M
of
M(M
-1)/2
independent standard-Gaussian random variables such that theran-dom measure
QM'
which is given by its (random) densityx f-7 exp{WM(X)} /
1M
exp{WM(X)} dx(2.5)
with (2.6) satisfies (2.7) for all ~ E D(N).GENTZ AND M. LOWE
(3.1) Remarks 2.2. 1. Note that the scaling factor NI/4 for the overlap vector is the
same as the one for the mean magnetization in the Curie-Weiss model at the
critical temperature, see [14, Theorem V.9.5]. Similar to that case (and, of
course, similar to the Hopfield model with a finite number of patterns) the distribution of the overlap is close to a non-Gaussian distribution.
2. Our condition M(N)15 IN --+ 0 on the growth rate of M is, of course,
em-barrassing. It is due to the simultaneous strong Gaussian approximation of
M(M -1)/2 variables. Any proof using the strong Gaussian approximation as
provided in [32], seems to produce conditions which are far off any reasonable
condition on the growth rate.
3. In fact, we are going to show that, under the conditions of the theorem,
ILM
f(x)QN(dx) -LM
f(X)QM(dx)1~
cN(Kf+
Ilflloo)
(2.8)
holds for all
f
E O(N) and all f E BL(]RM, ]R), where BL(]RM,]R) denotes theset of all bounded, Lipschitz continuous functions from]RM to]R, Kf denotes the
Lipschitz constant of f and
Ilflloo
= SUPxElRM If(x)l. This implies the theoremby (4.2) below.
3. STRONG GAUSSIAN ApPROXIMATION
In this section we are going to collect some facts about the so-called strong Gauss-ian approximation and apply them to the situation of our interest. The problem of the Gaussian approximation is quickly stated. Given a sequence (Xi)iEN of i.i.d. random vectors in ]Rd, we know that L~=lXi-scaled appropriately-converges in distribution to a Gaussian random vector Y. This vector can obviously be decom-posed again into a sum of "small" Gaussians. The question is now, whether we can also find Gaussian vectors
Yi
such that the difference"'(X, Y,n)
=
l~~dt
Xi -t
It
becomes small in a suitable sense.
This problem was first stated and treated in a one-dimensional setting by Koml6s, Major and Tusnady in [22]. The d-dimensional extension is due to Zaitsev [32] and Einmahl [12]. For a thorough treatment of the problem, we refer the reader to [33]. The form of the strong approximation we recall below proved useful in [10] and goes back to Einmahl and Mason [13].
Let PI andP2 be two probability measureson]Rd (endowed with the Borel a-field),
and for 8
>
0 let)..(PI,
g,
8)=
sup{PI(A) - P2(AO), P2(A) - PI (AO) : A C ]Rd closed}. (3.2) HereAO
=
{x E ]Rd : 3y E A such thatIlx -
yl12
~ 8} (3.3)is the closed 8-neighborhood of the set A.
Furthermore, let Xl, ... ,Xn be n E N independent random vectors in ]Rd with
JEXI = 0 and finite variance which satisfy the Bernstein-type condition
(3.5)
with some T for all m ~ 3 and all s,
t
E Rd.Under the condition (3.4), Zaitsev proved in [32, Theoreml.1] the following bound
on A(PI,n, P2,n, 8), where PI,n is the distribution of Xl
+ ... +
X n and P2,n is thed-dimensional normal distribution with mean zero and covariance matrix cov(XI )
+
... +
cov(Xn ) (see also [13]).Fact 3.1. For all n ~ 1 and all 8~ 0,
A(PI,n, P2,n,
8)
~ CI,d exp{-8/(C2,dT )}with CI,d
=
cld5/2 and C2,d=
c2d5/2 for numerical constants CI, C2>
0.As in [13], Fact 3.1, the following fact follows.
Fact 3.2. Let Xl, ... ,Xn be independent mean zero random vectors satisfying the
Bernstein-type condition (3.4). If the underlying probability space is rich enough,
then, for each 8 ~ 0, there exist independent Gaussian random vectors YI , ... ,Yn
with mean zero and
for alli E
{I, ...
,n}, such that(3.6)
where the constants CI,d, C2,d are the same as in Fact 3.1.
Corollary 3.3. In the situation of Fact 3.2, for each 8 ~ 0, there exists a mean
zero Gaussian random vectorY with covariance matrixcov(Y) = L~=lcov(Xi ) such
that
(3.7)
with the same constants CI,d, C2,d'
In our situation we want to apply Fact 3.2 and, in particular, Corollary 3.3 to the
M(M -
1)/2 dimensional vectors that contain the information of the mutual overlaps of the patterns in the ith component. More precisely, we will choose d=
M(M
-1)/2, n=
N,
andXi
=
(~f~nl::;Jl<v::;M in order to replace l/ylNL~1Xi
by a Gaussian random vectorTJ=
(TJJl,V )1::;Jl<v::;M. Observe that due to the independence of the ~f, we obtaincov(Xi )
=
Idfor each i, and hence also TJ will have identity covariance matrix. (By a slight abuse of notation, we denote the identity matrix by Id whatever the dimension of the underlying space
R
d is.) In order to apply Corollary 3.3, we have to check theBernstein-type condition (3.4). This is done in the following lemma.
Lemma 3.4. In the above setting Xl, ... ,X n fulfill the Bernstein-type condition
(3.4) with T = M.
B. GENTZ AND M. LOWE
Thus, for any choice ofs,t E JR.M and all m
2::
3IlE(s,
X i )2(t,x
i)m-21
:s;
Tm-21Itll;n-2
E(s, X i)2:s;
~m!Tm-2I1tll;n-2lE(s,
Xi?,where we have already chosen T
=
M. 0Now we are ready to deduce the desired approximation.
Corollary 3.5. If (O,:F,lP) is rich enough, for each Nand 8
2::
0, there exist amean zero Gaussian random variable TJ with covariance matrix Id and numerical
constants CI, C2
>
0, such thatProof. Apply Lemma 3.4 and Corollary 3.3 with T = M.
(3.8)
o
Remark 3.6. Observe that 8 in (3.8) may-and will indeed in our
applications-depend on Nand M.
4. PROOFS
To prove Theorem 2.1, we need to show that for large system size N the distribu-tion QN of the scaled overlap under the Gibbs measure (!N,l is close to the random
measure
Q
M with respect to the metric d on a set of large lP-measure. First we showthat QN and its smoothed version obtained by a Hubbard-Stratonovich transform
are close, so that we may investigate the Hubbard-Stratonovich transform instead of the measure itself. We recall the Hubbard-Stratonovich transform ofQN from [20].
The core of the proof is the investigation of the density of this Hubbard-Stratonovich transform by an adaptation of Laplace's method.
Notation 4.1. We denote by f..t
*
1/ the convolution of two measures f..t and 1/.Lemma 4.2. For all M
2::
8, all f E BL(JR.M ,JR.) and all probability measuresQ
onJR.M
,
If
f
d(Q
*
#(0, N-
Ij2
Id)) -
f
f
dQI
:s;
2V2K
r
la + Ilfllooe-
M,(4.1)
where Kf denotes again the Lipschitz constant of f and
Ilflloo
=
SUPxEJ~M If(x)I
asbefore. Now, with
go
=
g
n
{f : f(O)
=
O}
(4.2)
(4.3)
and
go
C BL(lRM ,JR.). Therefore, the following corollary is an immediateconse-quence of the preceding lemma.
Corollary 4.3. For all M
2::
8 and all probability measuresQ
on JR.M,d(Q*#(0,N-
IProof of Lemma4.2. Let f E BL(lRM ,lR) and let Q be an arbitrary probability
measure on lRM . Then, for 0
>
0,IJ f d(Q
*
N(O,N-1/2Id)) - J f dQI::; J J IB(O,6)(x)lf(x
+
y) - f(y)1 Q(dy) N(O, N-1/2Id)(dx)
(
vN)M/2
J
N+
211fll00 27r
IB(o,6)c(x)exp {-21Ixll~}
dx::; Kfo
+
21IfIl00l'M(B(0, oN1/2)C),
(4.5)whereI'M denotes the M-dimensional Gaussian measure with mean zero and the co-variance matrix being the identity matrix. The radiusTM satisfyingI'M(B(O, TM)) =
1/2is bounded by V2M for M
2::
8, cf. [18, Equation (4.4)]. Choosing 0= 2~,( ( 1/2
C)
1 { I[1/2 ] 2} 1 M ( )I'M B O,oN ) ::; "2 exp
-2
N O-TM ::; "2e- 4.6follows by [24, Theorem 1.2]. This concludes the proof. 0
The Hubbard-Stratonovich transform of the distribution of the scaled overlap is given by its density with respect to Lebesgue measure.
Lemma 4.4. Let
°
<
(l<
00 and a>
0. Then the convolution aXN,{3,a = QN
*
N(O, N(lId)(4.7)
(4.9)
x E lRM ,
of QN
=
{2N,{3( vamN)-1 with the M -dimensional Gaussian distribution with meanzero and covariance matrix ;(3Id is the random measure on lRM which is given by
the (random) density
f () - exp{-N(l<I>N,{3(X/va)} x E lRM ,
N,{3,a X -
J
IRM exp{-N(l<I>N,{3(x/va)}dx' (4.8)
with respect to the M-dimensional Lebesgue measure, where
1 1 N
<I>N,{3(X)
=
211xll~
-
(3N :L:)ogcosh((3(x,~i)),
i=1
depends on the random patterns. Here (., .) stands for the inner product in lRM .
We omit the proof as it follows by a straight-forward calculation similar to the ones given in [7, Lemma
2.2]
or [15, Lemma 3.3].Before turning to the proof of Theorem 2.1, we gather some estimates which will prove useful in the sequel. The first of these estimates is a bound on the operator norm of the random matrix arising from the patterns.
Lemma 4.5 ([6, Theorem 4.1]). There exist a constant K
>
°
and an N1 E N suchthat
(4.10) for all N
2::
N1 .10 B. GENTZ AND M. LOWE
For later use, we define
01(N) = {~: 111~~T~llop
-
(1+
va)2\
<
va}.
(4.11)In particular, we know that for N
2:
N1 ,~ E 01(N) and all x,y E jRM,I
~ ~
(x,
~i)(Y, ~i)
-
(x, y)
I
S;4y'(>lIxll,lIylJ,·
(4.12)
We also need the following estimates to treat terms which involve products of components ~f for four or six different values of f.-l. These are provided by the following lemma. For {j
>
0 let( u
{I~ t~:"~r'~r'~r'l
>
8y'(>}
J11, ... ,J14 t-1 UU
{I~ t~r'~r'~r'~r'~r'~r'l
>
8y'(>}
r
(4.13)
J11, ... ,J16 t=lwhere each of the unions is taken over all sets of pairwise different indices in
{I, ... ,M}.
Lemma 4.6. For every (j
>
0, there exists an N2(6) such that for all N2:
N2(6)lP{02(N,6)C} ::; exp{_62
M/4}.
(4.14)
Proof. Let
and
C
N.,(J1.1, ... , 1'6)
={I
~
t
~r'~r'~r'~r'~r'~r
I
>
8y'(>}.
(4.16)
For pairwise different indices f.-l1, ... , f.-l6 E {I, ...
,M},
Chebychev's inequality witht
=
6va
implieslP(BN,I5(f.-ll, ... , f.-l4)) ::; exp{-t6vaN} exp{Nt2/2} = exp{_62
M/2}
and, similarly,
Therefore,
lP(02(N,6)C)::;
(~M(M
-1)+
:!M(M -l)(M - 2)(M -
3))
exp{-{j2M/2}.(4.17)
Choosing
M
large concludes the proof. 0The next lemma provides a bound similar to (4.12) for terms involving the Gauss-ian fJ instead ofN-1/2~T~. Let
03(N, R,,.,;) =
{~
:II::
fJJ1,v(~)XJ1XVI
<
,.,;R2JMllxll~
Vx
E jRM }. (4.18) J1<V(4.23)
Lemma 4.7.
lP{n3(N,R,/~f}:S52Mexp{-K2R4M/16}.
Proof. Let x, Y E lRM . First note that LJ.l<v rJJ.l,VxJ.lYv can be viewed as the scalar
product of rJ and the vector (xJ.lYv)J.l<v and that
II(xJ.lYv)J.l<vI12
:S
2-
1/
21IxI121IyI12'
By Chebychev's inequality,lP{
L rJJ.l,VXJ.lYv2
K'}
<
exp{-tK'}
exp {t;
II(xJ.lYv)J.l<vll~}
J.l<V<
exp{-tK'}
exp {~ Ilxll~IIYII~}
(4.19)
fort
>
0.
Choosingt
=
2K'/(llxll~llyll~),ll'{~~p,vxpYv
2 ",}
:s
exp{
-IIXI~'I;YII1}
(4.20)
follows. To obtain a uniform bound, note that
lP{
3x E lRM : L rJJ.l,VXJ.lXV2
K'llxlI~}
=
lP{
3x EB(O,
1) :
L rJJ.l,VXJ.lXV2
K'}
J.l<V J.l<V
<
lP{
3x,Y EB(O,I):
LrJJ.l,vXJ.lYv2
K'}.
J.l<V
B(O,
1) being a (bounded) convex, balanced set in lRM , there exists a subsetD
c
B(O,
2)
such that B(O,1)
is contained in the convex hull of D and D has at most 5Melements (see for example
[31,
Lemma10.2
in the Appendix]). Now, by our previous bound and the definition of the set D,lP{
3x E lRM : L'TlJ.l,vXJ.lXV 2K'llxll~}
J.l<V
:S
lP{3X,y ED: LrJJ.l,vXJ.lYv2
K'}:S
52MxSUPDlP{L'TlJ.l,VXJ.lYv2
K'}
J.l<V ,yE J.l<V
:s
5'MX~~D
exp{
-lIxl~'I;YI11}
:s
5'Mexp{ -
~:}.
(4.21)
Choosing K'
=
KR2yIM
with K>°
concludes the proof. DWith these preparations we are able to prove Theorem 2.1.
Proof of Theorem 2.1. By
(4.2),
Theorem2.1
follows, once we have shown that,under the conditions of the theorem,
ILM
f(X)QN(dx) -LM
f(X)QM(dx)I
:S
cN(Kf+
Ilflloo)
(4.22)
holds for all~ E n(N) and all
f
E BL(lRM ,lR). By Lemma4.2,
we may replace QNby its Hubbard-Stratonovich transform.
So let
f
E BL(lRM ,lR). We need to investigatef
f(x) exp{-N~(x/Nl/4)}dxf
exp{-N~(x/Nl/4)}dx12
where
B. GENTZ AND M. LOWE
(4.24)
Consider the nominator first as the denominator is a special case of the nominator. The main contribution to the integral arises from the inner region
B(O,
RM1/4) andwe shall choose a suitable R
>
°
later on.In
the inner region as well as in theintermediate region
B(O,
rN1/4) \B(O,
RM1/4) with r>
°
to be chosen later, weinvestigate the behaviour of the integral in the nominator with the help of a Taylor expansion of <1>. The outer region
B(O,
r Nl/4Y is treated separately.Taylor expansion. Calculating the Taylor expansion of<1> around zero, we see that there exists a
e
E (0,1) such that<I'(x)
=
~
IIxlll-
~ ~ [~(X,~i)2
-
112
(x,
~i)4]
+
RN(x,
0,
(4.25) withwhere
h(t)
=
tan~~t))
[2 - sinh2(t)], t E Rcosh t
Regrouping the terms of the Taylor expansion of<1>, we find that
-N<1>(x/N1/4)
N
_ 1
II
114 1'"'*
2 2 1'"'*
1 '"' J1.I J1.2- -12 x 4 -
4
LJ XJ1.IXJ1.2+
2"
LJ XJ1.IXJ1.2VN
~~i ~i J1.1,J1.2 J1.1,J1.2 t=lN N
1
'"'*
3 1 '"'tJ1.I J1.2 1'"'*
2 1 '"'tJ1.I tJ1.2- 3
LJ XJ1.IXJ1.2N LJ"'i ~i -2"
LJ XJ1.IXJ1.2XJ1.3N LJ"'i "'iJ1.1,J1.2 i=l J1.I,J1.2 ,J1.3 i=l
(4.26)
(4.27)
(4.28)
(4.29)
l I N
12
2:*
XJ1.I X J1.2 X J1.3 X JJ4N2:~rl~r2~r3~r4
+
O(N/RN(x/N1/4,OI),
JJI,JJ2,J1.3,J1.4 i=lwhere
Ilxll~
=L:~1
xt·
Here and in the sequel, we use the notationL:*
J1.1, ••• ,J1.kfor summation over all k-tuples (J-l1,"" J-lk) E {I, ... ,
M}
with pairwise disjoint components.Let us consider the different ~-dependent terms. By the strong Gaussian approxi-mation Corollary 3.3, there exist a constant No EN and an M(M -1)/2-dimensional Gaussian vector TJ with mean zero and covariance matrix being the identity matrix
such that .
Oo(N,
ON)
={[I
~~w'm"<v
-
~112
<
ON }
with
for some K
>
0 satisfies(4.31)
for all N ~ No and
I
~
L'
x",x"'~t~t~r'
-
L
~""",x",x"'
I
<;
oMII(x", x",)",<",II,
<;
~lIxll~
Jll,Jl2 2=1 Jll<Jl2
(4.32)
for all ~ E Oo(N, bN).
The other ~-dependentterms become small due to the law of large numbers. For
N ~ N l and ~ E OleN), the bound (4.12) on the random matrix yields
(4.33)
as well as
(4.34)
Furthermore, for N ~ N 2
(b)
and ~ E 02(N,b),
by the definition of02(N,b),
11
1
2L'
x", x",x",x",
~
t
~t~r'~r'~r'l
(
4.35)Jll,Jl2,Jl3,Jl4 2=1
by!G. ""'
by!G.M
2 b(M
5) 1/2~ ~
L....J IX Jl1 XJl2XJl3XJl41~
12Ilxll~
=
12 NIlxll~·
It remains to consider the remainder of the Taylor expansion. Now,
Ih(t)1
~2\tl
and 0
< () <
1 together with Schwarz' inequality imply thatIRN(y,
~)I
<;
I:N
t(y,
~i)6
<;
:5L
Iy.,-·· y",11
~
t
~r'
...
~f'
I·
(4.36)2-1 Jll, ... ,Jl6 2-1
The right-hand side is bounded above by a combinatorial factor times the sum of terms similar to the ones treated above (with two, four or six different ~n plus the term arising from 1-"1 = ... = J.t6. This yields
[
(M5)
1/2
(M
7)1/2
]
IRN(y,~)1 ~
c
hllyll~
+
bN
Ilyll~
+
bN
lIyll~
+
lIyll~
for N ~ max{Nl ,N 2
(b)}
and ~ E Ol(N)n
02(N,b),
so thatNIRN(X/Nl/4,~)1~ ~ [hllxll~
+
2b(
M~/) 1/21Ixll~
+
Ilxll~].
(4.37)
(4.38)
l,From now on, we shall always assume that N ~ max{No, N l , N2
(b)}
and that14 B. GENTZ AND M. LOWE -Nif>(x/N1
/4) differs from
w(x) =
-1121Ixll~
-~
L
x~x~
+
L
7}JL,VxJLXVJL<V JL<V
by at most a constant times
(4.39)
[
M5
1/2] 1 [(M
7
)1/2]
IIxI16 gN(X)=
bMllxll~
+
va
+
b(
N )
Ilxll~
+
VN
va
+
8N
Ilxll~
+
vN·
(4.40)The inner region. For IIxl12 ~ RM1/4, the main contribution to 9N(X) arises from
the first summand. Therefore, we shall use the estimate
(M
I5) 1/2
9N(X) ~ hN(8, R)
=
N
(K+
8)R6 -+ 0, (4.41)provided MI5/N -+ O. (Recall that b
M = KM7
/VN.)
Therefore, the estimate forthe inner region is immediate: For j E BL(IRM,1R),
r
j(x)exp{-Nif>(x/N1/4)}dx } B(O,RMl/4)
= exp{O(hN(8,R)}
r
j(x)exp{w(x)}dx. (4.42)} B(O,RMl/4)
The intermediate region. For RMI/4 ~ IIxl12 ~ rN1/4,
which implies, that there exists an N3(8,r) E N such that
9N(X) ~ 8Mllxll~
+
2r211xll~(4.43)
(4.44)
for all N
2
N3(8, r), provided provided M7/N -+ O.AssumingN
2
max{No,N1 ,N2(b),N3(8} ande
E rlo(N,8N )nrl1(N) nrl2(N,8)n
rl3(N, R,
r;,)
from now on, our previous estimates together with the definition ofrl3(N, R,
r;,)
yield-Nif>(x/N1
/4) (4.45)
~ w(x)
+
O(gN(X))~ -1121Ixll~
-~
L
x~x~
+
L
7}JL,Vx JL XV+
O(8Mllxll~
+
2r21Ixll~)
JL<V JL<V
1 1
~ -121Ixll~
- 12[llxll~
- Il x 111]+
r;,R2JMllxll~
+
O(8Mllxll~
+
2r21Ixll~)·
For IIxl122
RM1/4, Ilxll~2
R2JMllxll~ is trivial. By choosing rand 0<
r;,
~ 1/48small enough, we see that there exists an N4(R, K) E N such that bM becomes so
small that
holds for all N
2:
N4(R, K) and all x from the intermediate region. Therefore, for allf
E BL(JRM ,JR) and N and ~ chosen as before,1
1
f(x) exp{ -Nif!(x/N1/4)} dxl {RMl/4:SllxI12:SrNl/4}~
Ilflloo1
exp { -~ VMllxll~}
dx {lIxIl2~RMl/4}~
Ilfllooexp{-R4M/48}r
exp {- R ZVMllxll~}
dxJIRM
48=
IIfliooexP{-R"M/48}(R;~)
MI2 (4.47) This bound will allow us to deduce that the integral over the intermediate region is negligible.The outer region. The investigation of the outer region consists of two parts. First, we show that there exists an ro
>
0 such that the integral overB(O,
r oN1/4y
is negligible and then, in a second step, we show that this ro can be replaced by an arbitrarily small r
>
o.
For convenience, we denote by fcw({3) the free energy in the Curie-Weiss model
at temperature 1/(3, i. e.,
fcw({3)
=
-%Z({3?+
logcosh({3z({3)).Then,
1
{I
} 1
log coshx ~ -{3XZ
+
max --{3t2+
log cosht
= -(3xz+
fcw(2{3),4 tEIR 4 4
which implies in particular that
VN
N-Nif!(x/N1/4) = --2-lIxll~
+
:?=logcosh(x/Nl/4'~i)z=l
(4.48)
(4.49)
< -
~llxll~
+
~ t(X'~i)Z
+
N
fcw(2).(4.50)
4yN i=l
Estimating the sum with the help of the bound (4.12) on the random matrix ~~T~, we see that there exist ro
>
0 and N52:
N1 such that-Nif!(x/N1/4)
~
-~llxll~
(4.51)holds for all x satisfying Ilxllz
2:
roN l/4,
all N2:
N5 and all ~ E Ol(N).Let now rNl/4 ~ Ilxllz ~ roNl/4 with an arbitrary r
E (0,
ro). First note thatif!(x/N1/4)
2:
JE{~(X/Nl/4,6)Z
-logcosh(x/N1/4,6)} (4.52)- sup
I
~
t
log cosh(x/N1/4,~i)
-
JElog cosh(x/N1/4,
6)I·
lIyl12:Sr o i=l16 B. GENTZ AND M. LOWE
The first summand on the right-hand side is bounded below by
cr,ro
=
inf E4>((Y,6)),y:r:::;IlyIl2~ro
where
(4.53)
4>(t)
=
t2/2
-logcosht, t E R, (4.54)attains its unique minimum at
t
=
O. The fact that (Y,6) is a (finite) Rademacher average (see [24, Chapter104], for instance), implies thatIP(I(Y, 6)1 ~ ~llyI12)
>
1/3 (4.55)(d. [17, Lemma 4.3]), so that
cr,ro= inf E4>((y,6)) >0, (4.56)
y:r~llyIl2~ro
because there is a set of positive IP-measure, on which 4> is bounded away from its unique minimum at zero.
The second summand on the right-hand side of (4.52) becomes small due to so-called self-averaging. Inspection of the proof of [17, Lemma 4.2] shows that not only
lim sup IN1
tf((X'~i))-Ef((X'~l))I=o
(4.57)N-+ooIIxl12~ro i=l
holds IP-almost surely for Lipschitz continuous
f,
but we obtained also bounds valid for large but fixed N:Lemma 4.8 ([17, Lemma 4.2]). There exist a constant c
>
0 and an N6 ~ N1 suchthat for all c
>
0 and all N ~ max{N6 ,2/c2}IP{ sup
I~ tlogCOSh(Y'~i)
-EIOgCOSh(Y,6)1~
(3+2ro)c}lIyl12~ro i=l
:S
2 exp{M(log(ro/c) +cn
exp{-Nc2/8} + IP(r21(N)C).
With
cr ro
c= '
2(3
+
2ro)and
r24(N,r,ro)
=
{~:
supI~ tlogCOSh(Y'~i)
-EIOgCOSh(Y,6)1:s cr;o} (4.58)lIyll2~ro i=l
we obtain the following corollary.
Corollary 4.9. There exist a constant K(r, ro)
>
0 and an N7(r, ro) E N such thatfor all N ~ N7(r, ro)
IP(r24(N, r, ro)C)
:S
exp{-K(r, ro)N} + IP(r21(Nt).Now, by our estimates on the two summands on the right-hand side of (4.52), we find
-NiP(x/N1/4)
:S
-Ncr,ro/2 (4.59)for all x such thatrN1
Gathering our estimates on the outer region yields
I
(
f(x)
exp{-NiIJ(x/N
1/4)}dxl
J{llxI12~rNl/4}
$ { Ilflloo exp {-
VNIIXII~}
dx
J{lIxIl2~roNl/4}
6+
1
Ilflloo exp{-NCr ,ro/2}dx
{rNl/4:SlIxll2:SroNl/4}
$llfII00[exp{-Nr5/12}+exp{-Ncr,ro/4}] (4.60)
for all
N
2:
N
8(r, ro)
for someN
8(r, ro)
E N.Completing the proof. l,From now on we shall always assume that
~ E
r!(N)
r!(N, R, r, ro, 0,1\,)
=
r!o(N,
ON)n
r!l(N)
n
r!2(N, 0)
n
r!3(N, R,
1\,)n
r!4(N, r, ro) (4.61)
and that
Note that there exists a constant L
>
0 such thatlP(r!(N)C)
$exp{-M/L},
(4.63)provided R is chosen large compared to I\, and M is large enough, d. Lemma 4.7. Naturally,
L
depends on our choice ofR, r, ro, {;
and 1\,.Let
f
E BL(JRM, JR) be arbitrary. We have already shown thatJ
f(x)
exp{-Nq,(x/N
1/4)}
dx
= exp{
O(hN(o, R)} (
f(x)
exp{w(x)}dx
J
B(O,RMI/4)+
0
(lIfll=ex
p {-R'M/48}
(R;~)
M!2)
+
0 (1lflloo [exp{-Nr5/12}
+ exp{-NCr,ro/4}])
(4.64)
with
hN(o,
R) ---+ O. Next, we want to replace the integral(
f(x)
exp{w(x)}dx
J
B(O,RMI/4)(4.65)
by the integral over JRM. First note, that (4.45) already provides an upper bound on
w(x),
valid for allx
satisfying IIxl122:
RM1/4:
1 1
R
2GENTZ AND M. LOWE (4.67) As an immediate consequence,
I
} {llxI122r
RM1 /4}f(x) exp{'lJ(x)} dxl
~
11J1100r
exp{ -~: VMllxll~}
dx
} {IIxI122RM1/4} ( 48 )M/2
~
Ilflloo exp{-R
4
M/48}
R2JM
'
which implies by (4.64) that
J
f(x)
exp{-NiJ>(x/N
1/4)} dx
=exp{O(h
N(8, R)}
r
f(x) exp{'lJ(x)} dx
J~M
( ( 48)M/2)
+
0 Ilfllooexp{-R
4
M/48}
R2JM
+
0 (1lflloo [exp{-Nr5/12}
+
exp{-Ncr,ro/4}] ).
(4.68)In order to compare
J~M
f(x)
exp{-NiJ>(x/N
1/
4)}
J~M exp{
-NiJ>(x/Nl/4)} dx
toJ~M
f(x) exp{'lJ(x)}
J~M
exp{'lJ(x)} dx '
we need a lower bound on J~M
exp{'lJ(x)} dx.
To obtain a lower bound on'lJ
first, we proceed as in (4.45):'lJ(x)
~
-1121Ixll:
- l [llxlli - Ilxll:] -f);R2VMllxll~ ~
-lllxlii -f);R2VMllxll~·
(4.69)For IIxl12 ~
RMI/4,
R
2'lJ(x)
~ -3VMllxll~
follows. (Recall, that f); ~ 1/48.) Now,
1
1
R2
1(3 )
M /2exp{'lJ(x)} dx
~
eXP{--VMllxll~}
dx
~
-
:ru
~M B(O,RM1/4) 3 . 2
R2
M
for
M
large enough, i. e.,N
~Ng(R)
for someNg(R)
EN.With these preparations, it is easy to see that
J~M
f(x)
exp{-NiJ>(x/N
1/
4)}
J~Mf(x) exp{'lJ(x)}
J~Mexp{
-NiJ>(x/Nl/4)} dx
J~Mexp{'lJ(x)} dx
~
IIfllooJ~M exp{'lJ~x)}
dx
+
0'
(4.70)
(4.71)
(4.73)
where we use 0 as an abbreviation for
o(
hN(8,R)
L
exp{w(x)}
dX)
+
0(exp{
-R'M/48}
(R:~
f')
+
0 (exp{-NT~/12}
+
exp{-NCr,ro/4}).By our lower bound on
flR
M exp{'1J(x)} dx, we see thatR
can be chosen so large that there exist a constant K>
0 and an NlO(R, T, TO, 6, K,) E N such thatflR
M f(x) exp{-N<I>(x/NI/4)}flR
M f(x) exp{'1J(x)}flR
M exp{-N<I>(x/NI/4)} dxflR
M exp{'1J(x)} dx:::; Ilflloo
[O(hN(6, R))+
O(exp{_R4M/K})]for all N ~ NlO(R, T, TO,6,K,). Now the theorem follows from Lemma 4.2 and
Lemma 4.4 with
O(N)
as defined in the beginning of this subsection andN ~ N N(R,T,To,6,K,)
max{No, NI ,N2(6), N3(6, T), N4(R,K), Ns, N6 ,N7(T, To), Ns(T, TO),
Ng(R), NlO(R, T, TO, 6, K,)}. (4.74)
o
REFERENCES
1. D.J. Amit, H. Gutfreund, and H. Sompolinsky, Statistical mechanics of neural networks near saturation, Ann. Phys. 173 (1987), 30 - 67.
2. R.N. Bhattacharya and R. Ranga Rao, Normal approximation and asymptotic expansion, Wiley, New York, 1976.
3. A. Bovier and V. Gayrard, An almost sure large deviation principle for the Hopfield model, Ann. Probab. 24 (1996), 1444-1475.
4. _ _ , An almost sure central limit theorem for the Hopfield model, Markov Processes Relat. Fields 3 (1997), 151-173.
5. _ _ , The retrieval phase of the Hopfield model: A rigorous analysis of the overlap distribu-tion, Probab. Theory Related Fields 107 (1997), 61-98.
6. _ _ , Hopfield models as generalized random mean field models, Mathematical Aspects of Spin Glasses and Neural Networks (A. Bovier and P. Picco, eds.), Progress in Probability, Birkhauser, Boston, 1998, pp. 3-89.
7. A. Bovier, V. Gayrard, and P. Picco, Gibbs states of the Hopfield model in the regime of perfect memory, Probab. Theory Related Fields 100 (1994),329-363.
8. _ _ , Gibbs states of the Hopfield model with extensively many patterns,J. Statist. Phys. 79 (1995), 395-414.
9. _ _ ,Large deviation principles for the Hopfield model and the Kac-Hopfield model,Probab. Theory Related Fields 101 (1995),511-546.
10. A. Bovier and D. Mason, Extreme value behaviour in the Hopfield model, preprint, preliminary version, 1998.
11. A. Bovier and P. Picco (eds.), Mathematical aspects of spin glasses and neural networks, Progress in Probability, Boston, Birkhiiuser, 1998.
12. U. Einmahl, Extensions of a result by Koml6s, Major, and Tusnady to the multidimensional case, J. Multivariate Anal. 28 (1989), 20-68.
13. U. Einmahl and D. Mason, Gaussian approximation of local empirical processes indexed by functions, Probab. Theory Related Fields 107(1997), 283-311.
14. R.S. Ellis, Entropy, large deviations, and statistical mechanics, Grundlehren der mathemati-schen Wissenschaften, vol. 271, Springer, New York, 1985.
B. GENTZ AND M. LOWE
15. R.S. Ellis and C. M. Newman, Limit theorems for sums of dependent random variables occur-ring in statistical mechanics, Z. Wahrscheinlichkeitstheorie verw. Gebiete 44 (1978), 117-139. 16. B. Gentz, An almost sure central limit theorem for the overlap parameters in the Hopfield
model, Stochastic Process. Appl. 62 (1996), 243-262.
17. ,A central limit theorem for the overlap in the Hopfield model,Ann. Probab. 24 (1996), 1809-1841.
18. ,A central limit theorem for the overlap in the Hopfield model,Ph.D. thesis, Universitat Zurich, Switzerland, 1996.
19. , On the central limit theorem for the overlap in the Hopfield model, Mathematical Aspects of Spin Glasses and Neural Networks (A. Bovier and P. Picco, eds.), Progress in Probability, Birkhauser, Boston, 1998, pp. 115-149.
20. B. Gentz and M. Lowe, The fluctuations of the overlap in the Hopfield model with finitely many patterns at the critical temperature,preprint, submitted, 1998.
21. J. J. Hopfield, Neural, networks and physical systems with emergent collective computational abilities, Proc. Natl. Acad. Sci. U.S.A. 79 (1982), 2554-2558.
22. J. Koml6s, P. Major, and G. Tusnady, An approximation of partial sums of independent RV's and the sample DF. I, Z. Wahrscheinlichkeitstheorie verw. Gebiete 32 (1975), 111-131. 23. J. Koml6s and R. Paturi, Convergence results in an associative memory model, Neural
Net-works 1 (1988), 239-250.
24. M. Ledoux and M. Talagrand, Probability in Banach spaces, Ergebnisse der Mathematik und ihrer Grenzgebiete, Springer, Berlin, 1991.
25. D. Loukianova, Lower bounds on the restitution error in the Hopfield model, Probab. Theory Related Fields 107 (1997),161-176.
26. M. Lowe, On the storage capacity of Hopfield models with weakly correlated patterns,to appear in Ann. Appl. Probab.
27. R.J. McEliece, E.C.Posner, E.R. Rodemich, and S. S. Venkatesh, The capacity of the Hopfield associative memory, IEEE Trans. Inform. Theory 33 (1987),461-482.
28. C. M. Newman, Memory capacity in neural network models: Rigorous lower bounds, Neural Networks 1 (1988), 223-238.
29. L. A. Pastur and A. L. Figotin, Exactly soluble model of a spin glass, SOy. J. Low Temp. Phys. 3 (1977), no. 6, 378-383.
30. , On the theory of disordered spin systems, Theor. Math. Phys. 35 (1977),403-414. 31. M. Talagrand, Rigorous results for the Hopfield model with many patterns, Probab. Theory
Related Fields 110 (1998), 177-276.
32. A. Yu. Zaitsev, On the Gaussian approximation of convolutions under multidimensional ana-logues of S.N. Bernstein inequality conditions, Probab. Theory Related Fields 74 (1987), 535-566.
33. A. Yu. Zaitsev, Multidimensional version of the results of Koml6s, Major, and Tusnady for vectors with finite exponential moments, Tech. Report 95-055, SFB 343, Bielefeld, 1995. (Barbara Gentz) WEIERSTRASS-INSTITUT FUR ANGEWANDTE ANALYSIS UND STOCHASTIK, MOHRENSTR. 39, D-10117 BERLIN, GERMANY
E-mail address.BarbaraGentz:gentz@wias-berlin.de
(Matthias Lowe) EURANDOM, PO Box 513, NL-5600 MB EINDHOVEN, THE NETHERLANDS E-mail address.MatthiasLowe:lowe@eurandom.tue.nl